A Two-Stage Modeling Framework for Analyzing COVID-19 Deaths By County

Farmer School of Business, Miami University

Farmer School of Business, Miami University

College of Public Health and Social Justice, Saint Louis University

1 R Setup and Required Packages

In this project, the open-source R programming language is used to model the progression in the COVID-19 pandemic in different U.S. counties. R is maintained by an international team of developers who make the language available at The Comprehensive R Archive Network. Readers interested in reusing our code and reproducing our results should have R installed locally on their machines. R can be installed on a number of different operating systems (see Windows, Mac, and Linux for the installation instructions for these systems). We also recommend using the RStudio interface for R. The reader can download RStudio for free by following the instructions at the link. For non-R users, we recommend the Hands-on Programming with R for a brief overview of the software’s functionality. Hereafter, we assume that the reader has an introductory understanding of the R programming language.

In the code chunk below, we load the packages used to support our analysis. Note that the code of this and any of the code chunks can be hidden by clicking on the ‘Hide’ button to facilitate the navigation. The reader can hide all code and/or download the RMD file associated with this document by clicking on the Code button on the top right corner of this document.

if(require(pacman)==FALSE) install.packages("pacman") # check to see if the pacman package is installed; if not install it
if(require(devtools)==FALSE) install.packages("devtools") # check to see if the devtools package is installed; if not install it

# to check and install if these packages are not found locally on machine
if(require(albersusa)==FALSE) devtools::install_github('hrbrmstr/albersusa') #install package if needed
if(require(albersusa)==FALSE) devtools::install_github('dreamRs/r2d3maps') #install package if needed


# check if packages are not installed; if yes, install missing packages
pacman::p_load(tidyverse, magrittr, janitor, dataPreparation, lubridate, skimr, # for data analysis
               COVID19, rvest, # for extracting relevant data
               DT, pander, stargazer, knitr, # for formatting and nicely printed outputs
               scales, RColorBrewer, DataExplorer, tiff, grid,# for plots
               plotly, albersusa, tigris, leaflet, tmap, # for maps
               zoo, fpp2, NbClust, # for TS analysis and clustering
               VIM, nnet, caret, # explanatory modeling
               conflicted) # for managing conflicts in functions with same names

# Handling conflicting function names from packages
conflict_prefer('combine', 'dplyr') # Preferring dplyr::combine over any other package
conflict_prefer('select', "dplyr") # Preferring dplyr::select over any other package
conflict_prefer("summarize", "dplyr") # Preferring dplyr::summarize over any other package
conflict_prefer("filter", "dplyr") # Preferring filter from dplyr
conflict_prefer("dist", "stats") # Preferring dist from stats
conflict_prefer("as.dist", "stats") # Preferring as.dist from stats

# Custom Functions
source('custom_functions.R')

# Setting the seed
set.seed(2020) # to assist with reproducibility
sInfo = sessionInfo() # saving all the packages/functions and session info

2 Extracting the Datasets

For our analysis, we fuse data from multiple sources. We describe the process of obtaining and merging each of these sources in the subsections below.

2.1 Time Series Data

In this section, we utilize the COVID19 package to obtain the following information: (Guidotti & Ardia, 2020)
- Confirmed cases, recoveries and deaths;
- Policy information (e.g., transport closing, school closing, closing event, movement restrictions, testing policies, and contact tracing); and
- Population and standard geographic information for each county.

From this information, we have also computed the new daily and weekly confirmed cases/deaths per county. The data is stored in a tidy format, but can be expanded to a wide format using pivot_wider() from the tidyverse package.

endDate = '2020-11-07'
endDatePrintV = format(ymd(endDate), format = "%b %d, %Y")

counties = covid19(country = "US", 
                   level = 3, # for county
                   start = "2020-03-01", # First Sunday in March
                   end = endDate, # end Date 
                   raw = FALSE, # to ensure that all counties have the same grid of dates
                   amr = NULL, # we are not using the apple mobility data for our analysis
                   gmr = NULL, # we are not using the Google mobility data for our analysis
                   wb = NULL, # world bank data not helpful for county level analysis
                   verbose = FALSE)

counties %<>% # next line removes non-contiguous US states/territories
  filter(!administrative_area_level_2 %in% c('Alaska', 'Hawaii', 'Puerto Rico', 'Northern Mariana Islands', 'Virgin Islands')) %>% 
  fastFilterVariables(verbose = FALSE) %>% #dropping invariant columns or bijections
  filter(!is.na(key_numeric)) %>%  # these are not counties
  group_by(id) %>% # grouping the data by the id column to make computations correct
  arrange(id, date) %>% # to ensure correct calculations
  mutate(day = wday(date, label = TRUE) %>% factor(ordered = F), # day of week
         newCases = c(NA, diff(confirmed)), # computing new daily cases per county
         newDeaths = c(NA, diff(deaths)) )  # computing new daily deaths per county

# manually identifying factor variables
factorVars = c("school_closing", "workplace_closing", "cancel_events",
               "gatherings_restrictions", "transport_closing", "stay_home_restrictions",
               "internal_movement_restrictions", "international_movement_restrictions",
               "information_campaigns", "testing_policy", "contact_tracing")

counties %<>% # converting those variables into character and then factor
  mutate_at(.vars = vars(any_of(factorVars)), .funs = as.character) %>% 
  mutate_at(.vars = vars(any_of(factorVars)), .funs = as.factor)

At this stage, we have only read the data based on the covid package. The resulting data is stored at an object titled counties, which contains 783,216 observations and 19 variables. Note that we have filtered observations that do not have a numeric key and removed some columns that do not add any value to future analysis (e.g., invariant cols).

2.2 Cross Sectional Data

In this analysis, we have extracted the following datasets:

1. The CDC’s Social Vulnerability Index, where we would utilize the following four summary theme ranking variables:
   a. Socioeconomic – RPL_THEME1
   b. Household Composition & Disability – RPL_THEME2
   c. Minority Status & Language – RPL_THEME3
   d. Housing Type & Transportation – RPL_THEME4
   Note that each of these indices were computed by the Centers for Disease Control and Prevention/ Agency for Toxic Substances and Disease Registry/ Geospatial Research, Analysis, and Services Program. The csv file titled: SVI2018_US_COUNTY.csv was downloaded on Nov 02, 2020 (click here to download the file after selecting counties for the Geography Type and CSV File for the File type and clicking GO).

2. We also extracted the following political and administrative variables, as of the beginning of the pandemic:
   a. Voting results for all counties in the 2016 Presidential elections: Based on MIT Election Data and Science Lab (2018), we have obtained the 2016 elections data by county, and used the data to compute the percentage of total votes that went to President Trump, with the underlying hypothesis that the politicization of COVID response (e.g., perception/willingness to use face masks, policies and the population’s reaction to the disease) may be explained by party affiliation.
   b. State Governor as of the beginning of the pandemic: Based on Wikipedia’s Table of State Governors (click for permanent link to the version we scraped), we scraped the Governor’s party affiliation since we hypothesized that it may impact the type of policies used on a state-level. Given that the District of Columbia does not have a governor, we imputed its value with “Democratic” since D.C.’s Mayor is a Democrat (and DC is not a state).
   c. State’s CDC Region Classification: We have also engineered a region variable based on the CDC’s 10 Regions Framework. While geographic regions are hypothesized to be a factor in disease outbreaks, we chose to utilize the CDC regions specifically based on the following explanation from the aforementioned link: “CDC’s National Center for Chronic Disease Prevention and Health Promotion (NCCDPHP) is strengthening the consistency and quality of the guidance, communications, and technical assistance provided to states to improve coordination across our state programs”

# (1) Obtaining the svi data
svi <- read_csv("../Data/Input/SVI2018_US_COUNTY.csv") %>% 
  clean_names() %>% # make everything lower_case
  select(location, fips, state, county, area_sqmi, e_totpop,
         rpl_theme1, rpl_theme2, rpl_theme3, rpl_theme4) %>% # columns of interest
  mutate(e_popdensity = e_totpop/ area_sqmi) %>% # computing population density
  filter(across(where(is.numeric), ~. >= 0)) %>%  # because NA's are coded as -999
  filter(!state %in% c('HAWAII', 'ALASKA')) # filtering to contiguous US


# (2) Political and Adminstrative Data
#### (a) 2016 Elections results on a county level
elections2016 <- read_csv("../Data/Input/countypres_2000-2016.csv") %>% # reading data
  clean_names() %>% 
  filter(year == 2016 & party == "republican" & # data for 2016 elections as %republican votes
         !state %in% c('Alaska', 'Hawaii') ) %>% # contiguous US states only
  mutate(fips = str_pad(fips, width = 5, side = 'left', pad = '0'), # proper FIPS data
         percRepVotes = 100*(candidatevotes/totalvotes) ) %>% # computing % republican votes (from total votes)
  select(fips, percRepVotes)

#### (b) State Governor
stateGovernor <- "https://en.wikipedia.org/w/index.php?title=List_of_United_States_governors&oldid=977828843" %>%
  read_html() %>% html_node("table:nth-child(9)") %>% html_table(header = 2, fill = TRUE) # scraping data
colnames(stateGovernor) <- stateGovernor[1,] %>% tolower() # lower case first row and making it colnames
stateGovernor <- stateGovernor[-1, c(1,5)] # dropping first row
stateGovernor$party %<>% recode(`Democratic–Farmer–Labor` = 'Democratic', #replacing  w/ Democratic
                               `Republican[note 1]` = 'Republican' ) #from Republican[note 1] to Republican
stateGovernor$state %<>%  toupper() # converting to upper case to match the names in the svi dataset
stateGovernor[51,] = c('DISTRICT OF COLUMBIA', 'Democratic') # since DC Mayor is a Democrat (DC is not a state)

#### (c) State Region Classification
cdcRegions = data.frame(state = c('Connecticut', 'Maine', 'Massachusetts', 'New Hampshire', 'Rhode Island' , 
                                   'Vermont', 'New York', # End of Region A
                                   'Delaware', 'District of Columbia', 'Maryland', 'Pennsylvania',
                                   'Virginia', 'West Virginia', 'New Jersey', # End of Region B
                                   'North Carolina', 'South Carolina', 'Georgia', 'Florida', # Region C
                                   'Kentucky', 'Tennessee', 'Alabama', 'Mississippi', # Region D
                                   'Illinois', 'Indiana', 'Michigan', 'Minnesota', 'Ohio',
                                   'Wisconsin', # End of Region E
                                   'Arkansas', 'Louisiana', 'New Mexico', 'Oklahoma', 'Texas', # Region F
                                   'Iowa', 'Kansas', 'Missouri', 'Nebraska', # Region G
                                   'Colorado', 'Montana', 'North Dakota', 'South Dakota',
                                   'Utah', 'Wyoming', # End of Region H
                                   'Arizona', 'California', 'Hawaii', 'Nevada', # Region I
                                   'Alaska', 'Idaho', 'Oregon', 'Washington' # Region J
                                   ) %>% toupper(),
                        region = c(rep('A', 7), rep('B', 7), rep('C', 4),
                                    rep('D', 4), rep('E', 6), rep('F', 5),
                                    rep('G', 4), rep('H', 6), rep('I', 4),
                                    rep('J', 4) ) )


# (3) Combining all the potential predictors in a crossSectionDatal Frame
crossSectionalData <- inner_join(svi, elections2016, by = 'fips') %>% 
  left_join(stateGovernor, by = 'state') %>% 
  left_join(cdcRegions, by = 'state')

# saving the results as a RDS File
saveRDS(crossSectionalData, '../Data/Output/crossSectionalData.rds')

# Tabulating the results and providing a way to export the table to different formats
datatable(crossSectionalData %>% select(-c(state, county)),
          extensions = c('FixedColumns', 'Buttons'), options = list(
            dom = 'Bfrtip',
            scrollX = TRUE,
            buttons = c('copy', 'csv', 'excel', 'pdf'),
            fixedColumns = list(leftColumns = 2))
          ) %>% 
  formatRound(columns= c('area_sqmi', 'rpl_theme1', 'rpl_theme2',
                         'rpl_theme3', 'rpl_theme4', 'e_popdensity', 'percRepVotes'),
              digits=2)

2.3 Exploratory Analysis

In this section, we perform an exploratory analysis on the data obtained from the multiple sources.

2.3.1 Cumulative Cases

noGoogleNAs <- filter(counties, !is.na(key_google_mobility)) # removing NAs from key_google_mobility
idIndex <- sample(noGoogleNAs$id, 9) # sampling 9 counties by id

# Saving cumulative deaths figure to an tiff file
tiff(filename = '../Figures/sampleCumulativeCases.tiff',
    width = 1366, height =768, pointsize = 16)
counties %>% filter(id %in% idIndex) %>% 
  ggplot(aes(x = date, y = confirmed, group = id, color = key_google_mobility)) +
  geom_line(size = 1.25) +
  scale_x_date(date_breaks = "1 month", date_labels = "%b") +
  facet_wrap(~ key_google_mobility, scales = 'free_y', ncol = 3) +
  theme(legend.position = 'none') + 
  labs(color = '', x = 'Month', y = 'Cumulative Cases By County') +
  scale_color_brewer(type = 'qual', palette = 'Paired')
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Creating an interactive plot for the markdown
p <- ggplot2::last_plot() + geom_line(size = 0.75) + # modifying the plot for plotly
  theme_bw(base_size = 9) + theme(legend.position = 'none') # to make margins smaller
ggplotly(p, height = 768) %>%  layout_ggplotly()

2.3.2 Cumulative Deaths

# Saving cumulative deaths figure to an tiff file
tiff(filename = '../Figures/sampleCumulativeDeaths.tiff',
    width = 1366, height =768, pointsize = 16)
counties %>% filter(id %in% idIndex) %>% 
  ggplot(aes(x = date, y = deaths, group = id, color = key_google_mobility)) +
  geom_line(size = 1.25) +
  scale_x_date(date_breaks = "1 month", date_labels = "%b") +
  facet_wrap(~ key_google_mobility, scales = 'free_y', ncol = 3) +
  theme(legend.position = 'none') + 
  labs(color = '', x = 'Month', y = 'Cumulative Deaths By County') +
  scale_color_brewer(type = 'qual', palette = 'Paired')
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Creating an interactive plot for the markdown
p <- ggplot2::last_plot() + geom_line(size = 0.75) + # modifying the plot for plotly
  theme_bw(base_size = 9) + theme(legend.position = 'none') # to make margins smaller
ggplotly(p, height = 768) %>%  layout_ggplotly()

2.3.3 RPL Theme 1

# Retrieving the U.S. county composite map as a simplefeature
cty_sf <- counties_sf("longlat") %>% filter(!state %in% c('Alaska', 'Hawaii')) # from albersua
cty_sf %<>% geo_join(crossSectionalData, by_sp= 'fips', by_df= 'fips')

# Saving a static version of the figure as tiff
tiff(filename = '../Figures/rplTheme1Map.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(cty_sf) + tm_polygons('rpl_theme1', title = 'RPL Theme 1: Socioeconomic', palette = "-Greens")
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/rplTheme1Map.tiff") %>% grid.raster()

2.3.4 RPL Theme 2

# Saving a static version of the figure (capitalizing on the tmap package)
tiff(filename = '../Figures/rplTheme2Map.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(cty_sf) + tm_polygons('rpl_theme2', title = 'RPL Theme 2: Household Composition & Disability', palette = "-Greens")
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/rplTheme2Map.tiff") %>% grid.raster()

2.3.5 RPL Theme 3

# Saving a static version of the figure (capitalizing on the tmap package)
tiff(filename = '../Figures/rplTheme3Map.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(cty_sf) + tm_polygons('rpl_theme3', title = 'RPL Theme 3: Minority Status & Language', palette = "-Greens")
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/rplTheme3Map.tiff") %>% grid.raster()

2.3.6 RPL Theme 4

# Saving a static version of the figure (capitalizing on the tmap package)
tiff(filename = '../Figures/rplTheme4Map.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(cty_sf) + tm_polygons('rpl_theme4', title = 'RPL Theme 4: Housing Type & Transportation', palette = "-Greens")
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/rplTheme4Map.tiff") %>% grid.raster()

2.3.7 Percent Republican Vote

# Saving a static version of the figure (capitalizing on the tmap package)
tiff(filename = '../Figures/percRepVotesMap.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(cty_sf) + tm_polygons('percRepVotes', title ='% Republican Votes', style = 'cont', palette = "div") +
  tm_layout(aes.palette = list(div = c('#2166AC', '#F7F7F7', '#B2182B')))
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/percRepVotesMap.tiff") %>% grid.raster()

2.3.8 CDC Regions

crossSectionalData$state %<>% tolower() %>% tools::toTitleCase()
# Retrieving the U.S. state composite map as a simplefeature
state_sf <- usa_sf("longlat") %>% filter(!name %in% c('Alaska', 'Hawaii')) # from albersua
state_sf %<>% geo_join(crossSectionalData, by_sp= 'name', by_df= 'state')

# Saving a static version of the figure (capitalizing on the tmap package)
tiff(filename = '../Figures/cdcMap.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(state_sf) + tm_polygons('region', title = 'CDC Region', palette = "Paired")
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/cdcMap.tiff") %>% grid.raster()

2.3.9 Governor’s Party Prior to November Elections

# Saving a static version of the figure (capitalizing on the tmap package)
tiff(filename = '../Figures/governorsMap.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(state_sf) + tm_polygons('party', title = "Governor's Party", palette = "div") +
  tm_layout(aes.palette = list(div = list('Democratic' = '#1F78B4', 'Republican' = '#E31A1C')))
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/governorsMap.tiff") %>% grid.raster()

3 Time-Series Clustering

It is important to note that, in our estimation, there are three important decisions to be made when performing time-series clustering:

• Preparation of the Different Time-Series to be Clustered In this section, we have (a) selected the new daily deaths per county as the primary variable of interest, (b) smoothed that variable using a seven-day moving average, and (c) scaled the observations within each county’s 7-day MA of new daily deaths such that it is bounded between 0 and 1. This allows us to compare the shape of the time-series/profile across counties of different populations and where the magnitude of the deaths is quite different.

• Choice of Distance Measure: The Euclidean distance, The Euclidean Distance i.e., the \(l_2\) norm, is the most commonly used distance measure since it is computationally efficient. However, it may not be suitable for applications where the time-series are of different length in addition to being sensitive to noise, scale and time shifts (Sardá-Espinosa, 2017).

• Choice of Clustering Algorithm: A large number of clustering algorithms have been proposed in the literature. Most common clustering approaches are shape-based, which include \(k-\)means clustering and hierarchical clustering. The reader is referred to Aghabozorgi et al. (2015) for a detailed review. In our preliminary analysis, we have chosen to use the hierarchical clustering approach since it provides an easy to understand dendogram and the number of counties was small. However, in our full analysis, we will use the \(k-\)means clustering algorithm since it is computationally efficient. Furthermore, we overcame the traditional limitation of having to pre-specify \(k\) by utilizing 26 indexes for determining the optimal number of clusters in a data set based on the excellent approach and package implementation of Charrad et al. (2014).

3.1 Data Preparation

clusteringPrep <- counties %>% # from the counties
  select(id, date, key_google_mobility, newDeaths) %>% # selecting minimal amount of cols for visual inspection
  arrange(id, date) %>% # arranged to ensure correct calculations
  mutate(newMA7 = rollmeanr(newDeaths, k = 7, fill = NA), # 7-day ma of new (adjusted) deaths
         maxMA7 = max(newMA7, na.rm = T), # obtaining the max per county to scale data
         scaledNewMA7 = pmax(0, newMA7/maxMA7, na.rm = TRUE) ) %>% # scaling data to a 0-1 scale by county
  select(id, key_google_mobility, date, scaledNewMA7) %>% # dropping the variable newDeaths
  pivot_wider(names_from = date, values_from = scaledNewMA7) # converting the data to a wide format for clustering

constantColumns  <- whichAreConstant(clusteringPrep, verbose = F) # identifying constant columns
datesDropped <- colnames(clusteringPrep)[constantColumns] # used for printing the names after the code chunk

clusteringPrep %<>% select(-all_of(constantColumns) ) %>%  # speeds up clustering by dec length of series
  as.data.frame() # data needs to be data frame for clustering
row.names(clusteringPrep) = clusteringPrep[,1] # needed for tsclust
clusteringPrep = clusteringPrep[,-1] # dropping the id column since it is now row.name

The following dates were removed from our data frame since the scaledNEWMA7 variable was constant across all counties: 2020-03-01, 2020-03-02, 2020-03-03, 2020-03-04, 2020-03-05, 2020-03-06 and 2020-03-07.

3.2 Clustering Contiguous U.S. Counties

clusteringPrep %<>% select(-c(key_google_mobility))  # removing this variable so we can cluster

nc  <- NbClust(clusteringPrep, distance = "euclidean", # euclidean distance
             min.nc = 2, max.nc = 50, # searching for optimal k between k=2 and k=50
             method = "kmeans", # using the k-means method
             index = "all") # using 26 of the 30 indices in the package

## *** : The Hubert index is a graphical method of determining the number of clusters.
##                 In the plot of Hubert index, we seek a significant knee that corresponds to a 
##                 significant increase of the value of the measure i.e the significant peak in Hubert
##                 index second differences plot. 
## 

## *** : The D index is a graphical method of determining the number of clusters. 
##                 In the plot of D index, we seek a significant knee (the significant peak in Dindex
##                 second differences plot) that corresponds to a significant increase of the value of
##                 the measure. 
##  
## ******************************************************************* 
## * Among all indices:                                                
## * 6 proposed 2 as the best number of clusters 
## * 12 proposed 3 as the best number of clusters 
## * 1 proposed 18 as the best number of clusters 
## * 1 proposed 23 as the best number of clusters 
## * 1 proposed 39 as the best number of clusters 
## * 1 proposed 45 as the best number of clusters 
## * 1 proposed 50 as the best number of clusters 
## 
##                    ***** Conclusion *****                            
##  
## * According to the majority rule, the best number of clusters is  3 
##  
##  
## *******************************************************************

kclus  <- nc$Best.partition %>% as.data.frame() %>% #obtaining the best partition/ cluster assignment for optimal k
  rename(., cluster_group = .) %>% rownames_to_column("County") 

#converting the wide to tall data and adding the cluster groupings
clusters  <- clusteringPrep %>% 
  rownames_to_column(var = "County") %>% 
  pivot_longer(cols = starts_with("2020"), names_to = "Date") %>% 
  inner_join(., kclus, by = "County") %>% 
  mutate(cluster_group = as.factor(cluster_group))

idClusters  <- clusters %>% select(c(County, cluster_group)) # creating a look-up table of county and cluster group
colnames(idClusters)  <- c('id', 'cluster_group') # renaming the columns
idClusters %<>%  unique() #removing the duplicates due to different dates (we had that to ensure that the clustering was applied correctly)

# Adding Cluster Grouping to a subset of the counties data frame
clusterCounties <- counties %>% 
  select(c(id, key_numeric, key_google_mobility, administrative_area_level_2, administrative_area_level_3)) %>% 
  inner_join(., idClusters, by ='id') %>% 
  mutate(cluster_group = paste0('C', cluster_group)) %>% 
  unique()

# saving the results as a RDS File
saveRDS(clusterCounties, '../Data/Output/clusterCounties.rds')

3.3 Visualizing the Clustering Results

In this subsection, we provide three plots:

• A paneled spaghetti plot, highlighting the median scaled time-series for profile for each cluster;
  
• A panel plot where the first, second and third quartiles of the scaled time-series for each cluster are compared; and
  
• An interactive chloropleth maps to visualize the spatial distribution of the clusters, where the reader can click on a given county to show: (a) county name, (b) assigned cluster, (c) population density, and (d) percentage of residents in poverty.

3.3.1 Spaghetti Plot

spaghettiDF <- counties %>% # from the counties
  select(id, date, newDeaths, key_google_mobility) %>% # selecting minimal columns
  left_join(clusterCounties[, c('id', 'cluster_group')], by = 'id') %>% # to get clusters
  arrange(id, date) %>% # arranged to ensure correct calculations
  mutate(newMA7 = rollmeanr(newDeaths, k = 7, fill = NA), # 7-day ma of new (adjusted) deaths
         maxMA7 = max(newMA7, na.rm = T), # obtaining the max per county to scale data
         scaledNewMA7 = pmax(0, newMA7/maxMA7, na.rm = TRUE) ) %>% 
  ungroup() %>% select(date, cluster_group, scaledNewMA7, key_google_mobility) %>% 
  group_by(date, cluster_group)

# Creating a Named Color Scale
colorPal <-  brewer.pal(n= levels(spaghettiDF$cluster_group) %>% length(), 'Set2')
names(colorPal) <- levels(spaghettiDF$cluster_group)

# Saving spaghetti plot to an tiff file
tiff(filename = '../Figures/spaghettiPlot.tiff', width = 1366, height =768, pointsize = 16)
spaghettiDF %>%  
  ggplot(aes(x = date, y = scaledNewMA7, color = cluster_group, group = key_google_mobility)) +
  scale_x_date(date_breaks = "1 month", date_labels = "%b") +
  geom_line(size = 0.25, alpha = 0.1) +
  stat_summary(aes(group = 1), 
               fun= median,
               geom = "line",
               size = 1.25, col = 'black') + 
  facet_wrap(~ cluster_group, ncol = 1) +
  theme(legend.position = 'none') + 
  labs(x = 'Month', y = 'Scaled New Deaths By Cluster By Day',
       caption = paste0('Solid black line represents the median for each cluster | 
       Based on Data from March 01, 2020 - ', endDatePrintV) )  +
  scale_color_manual(values = colorPal)
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/spaghettiPlot.tiff") %>% grid.raster()

3.3.2 Summary Plot

# Creating a data frame containing statistical summaries of the time series by cluster_group
summaryDf <- spaghettiDF %>% 
  summarise(Median = median(scaledNewMA7, na.rm= TRUE),
            `First Quartile` = quantile(scaledNewMA7, probs = 0.25, na.rm= TRUE),
            `Third Quartile` = quantile(scaledNewMA7, probs = 0.75, na.rm= TRUE)) %>% 
  pivot_longer(cols = c(`First Quartile`, Median, `Third Quartile`),
                        names_to = 'Statistic')

tiff(filename = '../Figures/summaryPlot.tiff', width = 1366, height =768, pointsize = 16)
summaryDf %>% 
  ggplot(aes(x = date, y = value, color = cluster_group, linetype =  Statistic)) +
  scale_x_date(date_breaks = "1 month", date_labels = "%b") +
  geom_line(size = 1.25) +
   scale_linetype_manual(values = c('dotted', 'solid', 'twodash')) +
  facet_wrap(~ cluster_group, ncol = 1) +
  theme(legend.position = 'top') + 
  labs(color = '', x = 'Month', y = 'Quartiles of Scaled New Deaths By Cluster By Day') +
  scale_color_manual(values = colorPal)
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/summaryPlot.tiff") %>% grid.raster()

3.3.3 Cluster Map

# Joining the clusterCounties results with the existing county simple features object (cty_sf)
clusterCounties$fips <- str_pad(clusterCounties$key_numeric, width = 5, side = 'left', pad = '0')
clusterCounties %<>% ungroup()
cty_sf %<>% left_join(clusterCounties[, c('fips', 'cluster_group')], by = 'fips') # adding cluster_group to cty_sf

# Creating a static visual for the paper
tiff(filename = '../Figures/clusterMap.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(cty_sf) + tm_polygons('cluster_group', title = 'Cluster #', palette = colorPal)
invisible( dev.off() ) # to suppress the unwanted output from dev.off


# Creating an interactive visual Using the Leaflet Package
#### Setting the Color Scheme
leafletPal <-  colorFactor('Set2', domain = cty_sf$cluster_group, na.color = "white")

#### The visual
leaflet(height=500) %>% # initializing the leaflet map
  setView(lng = -96, lat = 37.8, zoom = 3.8) %>% # setting the view on Continental US
  addTiles() %>% # adding the default tiles
  addPolygons(data = cty_sf, stroke = FALSE, fillColor = ~leafletPal(cty_sf$cluster_group), # adding the data
              weight = 2, opacity = 1, color = "white", dashArray = "3", fillOpacity = 0.7, # adding color specs
              popup = paste("County:", cty_sf$name, '<br>', 
                            "Cluster #:", cty_sf$cluster_group, '<br>',
                            "Population Density:", round(cty_sf$e_popdensity, 1), '<br>')) %>% #pop-up Menu
  addLegend(position = "bottomleft", pal = leafletPal, values =  cty_sf$cluster_group, 
            title = "Cluster #", opacity = 1) # legend formatting

4 Explanatory Modeling of Cluster Assignments

In the previous section, we showed that by using solely a scaled and smoothed time series of daily deaths per county, the counties are grouped into 3 categories (whose time-series have distinct shapes based on the Euclidean distance measure). In this section, we attempt to model the factors that are associated with the cluster assignment.

4.1 Descriptive Statistics

4.1.1 Skimming the Data

# Combining multiClass response with potential predictors
multiClassDF <- select(clusterCounties, fips, cluster_group) %>% 
  left_join(crossSectionalData, by = 'fips')  %>% 
  select(-c(e_totpop, area_sqmi, state, county))

saveRDS(multiClassDF, '../Data/Output/multiClassDF.rds') # saving the data

skim(multiClassDF) # printing a nice summary table of the data

Data summary

Name	multiClassDF
Number of rows	3108
Number of columns	11
_______________________
Column type frequency:
character	5
numeric	6
________________________
Group variables	None

Variable type: character

skim_variable	n_missing	complete_rate	min	max	empty	n_unique	whitespace
fips	0	1	5	5	0	3108	0
cluster_group	0	1	2	2	0	3	0
location	2	1	16	42	0	3106	0
party	2	1	10	10	0	2	0
region	2	1	1	1	0	10	0

Variable type: numeric

skim_variable	n_missing	complete_rate	mean	sd	p0	p25	p50	p75	p100	hist
rpl_theme1	2	1	0.50	0.29	0.00	0.25	0.50	0.75	1.00	▇▇▇▇▇
rpl_theme2	2	1	0.50	0.29	0.00	0.25	0.50	0.75	1.00	▇▇▇▇▇
rpl_theme3	2	1	0.50	0.29	0.00	0.25	0.50	0.75	1.00	▇▇▇▇▇
rpl_theme4	2	1	0.50	0.29	0.00	0.25	0.50	0.75	1.00	▇▇▇▇▇
e_popdensity	2	1	273.11	1813.57	0.15	17.37	45.26	118.68	72052.99	▇▁▁▁▁
percRepVotes	2	1	63.31	15.63	4.09	54.48	66.35	74.92	96.03	▁▁▅▇▃

4.1.2 Correlation Plot Among Continuous Predictors

tiff(filename = '../Figures/corrPlot.tiff',
    width = 1366, height =768, pointsize = 16)
na.omit(multiClassDF) %>% 
  plot_correlation(ggtheme = theme_bw(), type = 'c')  # compute corr among only continuous vars
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/corrPlot.tiff") %>% grid.raster()

4.1.3 Boxplot By Cluster

# Saving cumulative deaths figure to a tiff file
tiff(filename = '../Figures/boxPlot.tiff',
    width = 1366, height =768, pointsize = 16)
multiClassDF %>% 
  plot_boxplot(by = 'cluster_group', ncol = 2L, 
               ggtheme = theme_bw(),
               geom_boxplot_args = list('outlier.shape' = 1))
invisible( dev.off() ) # to suppress the unwanted output from dev.off

# Printing a png version of the plot in Markdown (lower quality image for quicker compilation of HTML)
readTIFF("../Figures/boxPlot.tiff") %>% grid.raster()

4.2 Explanatory Modeling Using Multinomial Regression

4.2.1 Model Building

df <- multiClassDF %>% select(-c(fips, location)) %>% 
  mutate(cluster_group = as.factor(cluster_group))

# setting the reference level to category with max frequency
df$clustReLeveled <-  relevel(df$cluster_group, ref = maxCat(df$cluster_group) )
df  <-  df %>% select(-c(cluster_group)) # removed since we created a reLeveled version and stored it in a diff va

finalModel <-  multinom(clustReLeveled ~ ., data = df) # create model

## # weights:  54 (34 variable)
## initial  value 3412.289769 
## iter  10 value 2279.964165
## iter  20 value 2165.730051
## iter  30 value 2135.056284
## iter  40 value 2118.518517
## final  value 2117.666623 
## converged

4.2.2 Resulting Model

# Saving the results as a latex table, but not printing it out in the Markdown document
invisible(stargazer(finalModel, type = 'latex', p.auto = FALSE, out="../Data/Output/multi.tex", 
                    single.row = TRUE, header = FALSE))

# tabulating model results as an HTML, which we print below
stargazer(finalModel, type = 'html', p.auto = FALSE, out="../Data/Output/multi.html", single.row = FALSE)

	Dependent variable:
	C1	C2
	(1)	(2)
rpl_theme1	-0.816***	-0.539***
	(0.090)	(0.100)
rpl_theme2	0.601***	0.287
	(0.188)	(0.182)
rpl_theme3	1.988***	2.493***
	(0.143)	(0.117)
rpl_theme4	1.106***	0.633***
	(0.149)	(0.142)
e_popdensity	0.003***	0.003***
	(0.0003)	(0.0003)
percRepVotes	-0.012***	-0.028***
	(0.003)	(0.003)
partyRepublican	1.064***	0.377***
	(0.121)	(0.132)
regionB	1.901***	-0.398**
	(0.184)	(0.171)
regionC	3.404***	-0.880***
	(0.127)	(0.175)
regionD	3.494***	-0.228
	(0.129)	(0.170)
regionE	1.715***	-0.142
	(0.154)	(0.128)
regionF	2.325***	-1.708***
	(0.111)	(0.159)
regionG	0.889***	-1.848***
	(0.190)	(0.207)
regionH	0.636***	-1.783***
	(0.235)	(0.222)
regionI	3.034***	-2.125***
	(0.191)	(0.077)
regionJ	2.136***	-1.716***
	(0.190)	(0.073)
Constant	-4.871***	-0.873***
	(0.078)	(0.089)
Akaike Inf. Crit.	4,303.333	4,303.333
Note:	p<0.1; p<0.05; p<0.01

4.2.3 Model’s Performance

# examining how well the model performed on our entire dataset
# Recall that we are fitting an explanatory model, and not a predictive model
predictedClass <- predict(finalModel, df)

# Computing the Confusion Metrics and By Class Metrics
confMatrix = confusionMatrix(predictedClass, df$clustReLeveled)
saveRDS(confMatrix, '../Data/Output/confMatrix.rds') # saving the data

# Printing the Resulting Tables Nicely
pander(confMatrix$table)

	C3	C1	C2
C3	1639	284	214
C1	153	474	94
C2	40	32	176

pander(confMatrix$byClass)

Table continues below

	Sensitivity	Specificity	Pos Pred Value	Neg Pred Value
Class: C3	0.8947	0.6091	0.767	0.8008
Class: C1	0.6	0.8934	0.6574	0.8675
Class: C2	0.3636	0.9725	0.7097	0.8922

Table continues below

	Precision	Recall	F1	Prevalence	Detection Rate
Class: C3	0.767	0.8947	0.8259	0.5898	0.5277
Class: C1	0.6574	0.6	0.6274	0.2543	0.1526
Class: C2	0.7097	0.3636	0.4809	0.1558	0.05666

	Detection Prevalence	Balanced Accuracy
Class: C3	0.688	0.7519
Class: C1	0.2321	0.7467
Class: C2	0.07985	0.6681

pander(confMatrix$overall)

Table continues below

Accuracy	Kappa	AccuracyLower	AccuracyUpper	AccuracyNull
0.737	0.4968	0.7211	0.7524	0.5898

AccuracyPValue	McnemarPValue
5.86e-66	1.016e-40

4.2.4 Visualizing the Model’s Predictions

predictedProbs <- fitted(finalModel) # computing predicted probabilities for each of the cluster outcome levels
mapResults <- cbind(na.omit(multiClassDF), predictedProbs) # col binding predProbs for Each Cluster with multiClassDF

# Finding indices to subset the data
numberOfClusters <- unique(mapResults$cluster_group) %>% as.character() %>% length() 
startCol <- ncol(mapResults) - numberOfClusters + 1
endCol <- ncol(mapResults)

# Finding whether the predicted and actual clusters matched for each county
mapResults$LargestProbCluster <- colnames(mapResults[, startCol:endCol])[apply(mapResults[, startCol:endCol], 1, which.max)] 
mapResults$match <- ifelse(mapResults$cluster_group == mapResults$LargestProbCluster, 'Yes', 'No') %>% as.factor()

# Retrieving the U.S. county composite map as a simplefeature (since it has been overwritten)
cty_sf <- counties_sf("longlat") %>% filter(!state %in% c('Alaska', 'Hawaii')) # from albersua
cty_sf %<>% geo_join(mapResults, by_sp= 'fips', by_df= 'fips')

# Creating a static visual for the paper
tiff(filename = '../Figures/clusterMatchMap.tiff', width = 1366, height =768, pointsize = 16)
tm_shape(cty_sf) + tm_polygons('match', title = 'Cluster Match', style = 'cont', palette = "div") +
  tm_layout(aes.palette = list(div = list("Yes" = "#CAB2D6", "No" = "#6A3D9A")))
invisible( dev.off() ) # to suppress the unwanted output from dev.off


# Creating an interactive visual Using the Leaflet Package
#### Setting the Color Scheme
leafletPal <-  colorFactor(palette = c("#CAB2D6", "#6A3D9A"), levels = c('Yes', 'No'), na.color = "white")

#### The visual
leaflet(height=500) %>% # initializing the leaflet map
  setView(lng = -96, lat = 37.8, zoom = 3.8) %>% # setting the view on Continental US
  addTiles() %>% # adding the default tiles
  addPolygons(data = cty_sf, stroke = FALSE, fillColor = ~leafletPal(cty_sf$match), # adding the data
              weight = 2, opacity = 1, color = "white", dashArray = "3", fillOpacity = 0.7, # adding color specs
              popup = paste("County:", cty_sf$name, '<br>', 
                            "Cluster #:", cty_sf$cluster_group, '<br>',
                            "Cluster Predicted:", cty_sf$LargestProbCluster, '<br>',
                            "Cluster Match:", cty_sf$match, '<br>')) %>% #pop-up Menu
  addLegend(position = "bottomleft", pal = leafletPal, values =  cty_sf$match, 
            title = "Cluster Match", opacity = 1) # legend formatting

5 Concluding Remarks

In this R Markdown document, we have shown that our proposed two stage framework for modeling the smoothed and scaled time series of new daily cases can provide insights into the shape of the outbreak’s time-series and some of its associated factors. Specifically, we have shown that:

• On a county-level, the time series of COVID-19 new daily cases can be clustered into 3 clusters.

• Using a multinomial regression model, we have shown/quantified the impact of the following factors: rpl_theme1, rpl_theme2, rpl_theme3, rpl_theme4, e_popdensity, percRepVotes, party and region on the odds of being in a specific cluster when compared to the baseline.

---

6 References

Aghabozorgi, S., Shirkhorshidi, A. S., & Wah, T. Y. (2015). Time-series clustering–a decade review. Information Systems, 53, 16–38.

Charrad, M., Ghazzali, N., Boiteau, V., & Niknafs, A. (2014). NbClust: An r package for determining the relevant number of clusters in a data set. Journal of Statistical Software, Articles, 61(6), 1–36. https://doi.org/10.18637/jss.v061.i06

Guidotti, E., & Ardia, D. (2020). COVID-19 data hub. Journal of Open Source Software, 5(51), 2376. https://doi.org/10.21105/joss.02376

MIT Election Data and Science Lab. (2018). County Presidential Election Returns 2000-2016 (Version V6) [Data set]. Harvard Dataverse. https://doi.org/10.7910/DVN/VOQCHQ

---

7 Appendix

In this appendix, we print all the R packages used in our analysis and their versions to assist with reproducing our results/analysis.

pander(sessionInfo(), compact = TRUE)

R version 4.0.3 (2020-10-10)

Platform: x86_64-w64-mingw32/x64 (64-bit)

locale: LC_COLLATE=English_United States.1252, LC_CTYPE=English_United States.1252, LC_MONETARY=English_United States.1252, LC_NUMERIC=C and LC_TIME=English_United States.1252

attached base packages: grid, stats, graphics, grDevices, utils, datasets, methods and base

other attached packages: sf(v.0.9-6), conflicted(v.1.0.4), caret(v.6.0-86), lattice(v.0.20-41), nnet(v.7.3-14), VIM(v.6.0.0), colorspace(v.1.4-1), NbClust(v.3.0), expsmooth(v.2.3), fma(v.2.4), forecast(v.8.13), fpp2(v.2.4), zoo(v.1.8-8), tmap(v.3.2), leaflet(v.2.0.3), tigris(v.1.0), plotly(v.4.9.2.1), tiff(v.0.1-5), DataExplorer(v.0.8.1), RColorBrewer(v.1.1-2), scales(v.1.1.1), knitr(v.1.30), stargazer(v.5.2.2), pander(v.0.6.3), DT(v.0.16), rvest(v.0.3.6), xml2(v.1.3.2), COVID19(v.2.3.1), skimr(v.2.1.2), dataPreparation(v.0.4.3), progress(v.1.2.2), Matrix(v.1.2-18), lubridate(v.1.7.9), janitor(v.2.0.1), magrittr(v.1.5), forcats(v.0.5.0), stringr(v.1.4.0), dplyr(v.1.0.2), purrr(v.0.3.4), readr(v.1.4.0), tidyr(v.1.1.2), tibble(v.3.0.4), tidyverse(v.1.3.0), albersusa(v.0.4.1), devtools(v.2.3.2), usethis(v.1.6.3), pacman(v.0.5.1) and ggplot2(v.3.3.2)

loaded via a namespace (and not attached): tidyselect(v.1.1.0), htmlwidgets(v.1.5.2), ranger(v.0.12.1), maptools(v.1.0-2), pROC(v.1.16.2), munsell(v.0.5.0), codetools(v.0.2-16), units(v.0.6-7), withr(v.2.3.0), highr(v.0.8), uuid(v.0.1-4), rstudioapi(v.0.11), stats4(v.4.0.3), robustbase(v.0.93-6), vcd(v.1.4-8), TTR(v.0.24.2), labeling(v.0.4.2), repr(v.1.1.0), farver(v.2.0.3), rprojroot(v.1.3-2), vctrs(v.0.3.4), generics(v.0.1.0), ipred(v.0.9-9), xfun(v.0.19), R6(v.2.5.0), assertthat(v.0.2.1), networkD3(v.0.4), rgeos(v.0.5-5), gtable(v.0.3.0), lwgeom(v.0.2-5), processx(v.3.4.4), timeDate(v.3043.102), rlang(v.0.4.8), splines(v.4.0.3), rgdal(v.1.5-18), lazyeval(v.0.2.2), ModelMetrics(v.1.2.2.2), selectr(v.0.4-2), dichromat(v.2.0-0), broom(v.0.7.2), reshape2(v.1.4.4), yaml(v.2.2.1), abind(v.1.4-5), modelr(v.0.1.8), crosstalk(v.1.1.0.1), backports(v.1.2.0), quantmod(v.0.4.17), lava(v.1.6.8), tools(v.4.0.3), ellipsis(v.0.3.1), raster(v.3.3-13), sessioninfo(v.1.1.1), Rcpp(v.1.0.5), plyr(v.1.8.6), base64enc(v.0.1-3), classInt(v.0.4-3), ps(v.1.4.0), prettyunits(v.1.1.1), rpart(v.4.1-15), fracdiff(v.1.5-1), tmaptools(v.3.1), haven(v.2.3.1), fs(v.1.5.0), leafem(v.0.1.3), data.table(v.1.13.2), openxlsx(v.4.2.3), lmtest(v.0.9-38), reprex(v.0.3.0), pkgload(v.1.1.0), hms(v.0.5.3), evaluate(v.0.14), XML(v.3.99-0.5), rio(v.0.5.16), readxl(v.1.3.1), gridExtra(v.2.3), testthat(v.3.0.0), compiler(v.4.0.3), KernSmooth(v.2.23-18), crayon(v.1.3.4), htmltools(v.0.5.0), DBI(v.1.1.0), dbplyr(v.1.4.4), MASS(v.7.3-53), rappdirs(v.0.3.1), boot(v.1.3-25), car(v.3.0-10), cli(v.2.1.0), quadprog(v.1.5-8), gower(v.0.2.2), parallel(v.4.0.3), igraph(v.1.2.6), pkgconfig(v.2.0.3), foreign(v.0.8-80), laeken(v.0.5.1), sp(v.1.4-4), recipes(v.0.1.14), foreach(v.1.5.1), prodlim(v.2019.11.13), snakecase(v.0.11.0), callr(v.3.5.1), digest(v.0.6.27), rmarkdown(v.2.5), cellranger(v.1.1.0), leafsync(v.0.1.0), curl(v.4.3), urca(v.1.3-0), lifecycle(v.0.2.0), nlme(v.3.1-150), jsonlite(v.1.7.1), tseries(v.0.10-47), carData(v.3.0-4), desc(v.1.2.0), viridisLite(v.0.3.0), fansi(v.0.4.1), pillar(v.1.4.6), survival(v.3.2-7), httr(v.1.4.2), DEoptimR(v.1.0-8), pkgbuild(v.1.1.0), glue(v.1.4.2), xts(v.0.12.1), remotes(v.2.2.0), zip(v.2.1.1), png(v.0.1-7), iterators(v.1.0.13), leaflet.providers(v.1.9.0), class(v.7.3-17), stringi(v.1.5.3), blob(v.1.2.1), stars(v.0.4-3), memoise(v.1.1.0) and e1071(v.1.7-4)

---

1. Email: fmegahed@miamioh.edu | Phone: +1-513-529-4185 | Website: Miami University Official↩︎

2. Email: farmerl2@miamioh.edu | Phone: +1-513-529-4823 | Website: Miami University Official↩︎

3. Email: steve.rigdon@slu.edu | Website: Saint Louis University Official↩︎